The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions in SKILL.md include a Python script executed via Bash that interpolates a user-supplied placeholder ([Brand_Name]) directly into a string literal. This creates a risk of Python code injection or shell command failure if the input contains single quotes or metacharacters.
[EXTERNAL_DOWNLOADS]: The analysis procedure performs network requests to retrieve data from well-known services.
Evidence: The script queries https://en.wikipedia.org/w/api.php and https://www.wikidata.org/w/api.php to verify brand entities.
Note: These interactions target high-authority official endpoints for verification and do not involve remote code execution.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface due to the processing of untrusted external content.
Ingestion points: Untrusted data enters the context from user input and platforms such as YouTube and Reddit via the WebFetch tool.
Boundary markers: The skill lacks instructions for the agent to use delimiters or ignore instructions embedded in the fetched platform data.
Capability inventory: The skill utilizes Bash (for execution), WebFetch (network access), and Write (to save reports).
Sanitization: No evidence of sanitization for platform-derived content is provided before its inclusion in generated reports.

geo-brand-mentions